Don't forget we are talking about CPUs that are getting close to Pentium 3 class performance. Intel proved back in the Pentium vs PowerPC days that x86 can compete well against superior architectures. In this fight they have a lot more performance to work with.
I do agree with your CPU-becoming-commoditized point, but Intel is very aware of that (cite: how they keep Atom performance just enough higher than ARM, but a lot slower than their more profitable higher end chips). It's a difficult area, but Intel is aware of the balancing act they have to do. I think their strategy is to increase performance of non-CPU components of their chipsets (ie, make sure Atom kills ARM on I/O) in order to keep their lead in the datacenter.
PowerPC owns gaming. MIPS seems to still be common in networking devices. ARM is mobile and the rest is Intel and noise.
Seems to me that the killer feature ARM has isn't technical at all, it's that you can buy a core, graft on a DSP if your core doesn't have one (OMAPs already do) and your secret sauce needs a DSP (half the phone vendors have their own audio enhancement stuff or echo cancellation code that they think makes the difference) graft on a mobile chip and fab your own part. Then you add memory, flash, a battery and you've pretty much got a phone.
I'm not aware of any medium volume products that use custom Intel based hardware, I have no idea what the costs or terms are but I'm under the impression that unless you're going to make 5million chips, it's just not worth it with Intel parts. Now say intel opensourced an SSE echo killer and built some kind of 2watt Atom with GSM built in style chipset and it cost $15, I bet you'd start thinking ARM was in trouble.
Intel can pretty much build the highest performing per dollar chips, they can build them cheaper than anyone else, they have much more reliable processes than anyone else and they've shown time and time again that when they put their mind to it they can compete with anyone else.
I'd need to see some really compelling evidence that ARM was moving in to the serving market in any meaningful way. I thought they had a match with PowerPC, thought so with AMD, thought Linux was going to make alternative platforms viable, and probably thought they were in trouble several other times but they are fiendishly good.
True - see also this:
http://codingrelic.geekhold.com/2010/08/x86-vs-arm-mobile-cp...
However, if x86 has neither a process or a compatibility advantage, then even a tiny technical disadvantage turns into a tiny extra cost, which can be important for high volume chips.
Not true.
The marginal cost of pretty much anything on a chip itself is close to zero. For example, most 3 core chips are actually 4 core chips with one disabled. Putting the extra silicon on the chip is effectively free.
The money goes in the investment in the factory and the R&D, NOT the raw materials or production costs.
If you need that silicon to actually work, then it's not free. Floor-sweeping is necessary because the more stuff you put on a chip, the more likely it is to have defects. That really is an extra cost. I don't think there is any way around that.
But I think you are overestimating how low-end these chips are. See http://arstechnica.com/gadgets/news/2008/02/small-wonder-ins... for example, which shows that the Silverthorne architecture (1st Gen Atom) has pretty much the same transistor count as a Pentium 4.
No one thought that the fact the Pentium 4 had to support x86 was a significant factor vs other architectures. For the chips we are talking about, the cache memory takes more transistors, so the quality thing is pretty insignificant too.
The situation is even worse for graphics cards
Also GPUs have their uses, but general-purpose they are not. GPUs are complements to CPUs (just as coprocessors specialized for floating-point calculations where back in the day).
Most people here probably haven't lived the day when 80387 had to be installed alongside i386 if you wanted decent graphics performance. It is actually interesting that it takes so long for GPUs to be merged with X86.
In my opinion ARM is the only credible threat in the data-center, especially for servers that are mostly I/O bound. But on the other hand costs can skyrocket when scaling horizontally and having 500 HTTP servers (like Digg) is not at all fun.
And IMHO, x86 chips have better performance / watt.
For desktop computing X86 will still dominate, at least for the following 10 years, simply because of Windows; which runs on over 1 billion computers.
Yeah, I know it is fashionable on HN to say Windows is not relevant anymore. Doesn't make it true.
for some tasks(like HPC , multimedia) , the gpu can handle most of the computation , and the cpu is somewhat of a sidekick(like in the nVidia ION netbooks , for example).
while windows would still dominate, the importance of a strong x86 cpu would decrease, by all the changes in the ecosystems , and the costs of the x86 processor will decrease.
I had a phone conversation with one of their managers about a year ago (I was trying to see could I get a sample PCIe board) and they are damned expensive, so it would only make sense if you have something that 1) parallelizes very well, 2) actually needs the performance, 3) runs on Linux* or can be easily ported to their architecture and 4) you have high enough volume to make their overheads worthwhile. On the plus side, they seemed were very willing to lend their engineering team to help port to their architecture (as part of the devkit cost). Sadly, they don't have evaluation boards - you pretty much have to buy a devkit to evaluate it and the devkits costs approx. 5 times the price of the lowest end PCIe board (I'm not sure what hardware this included and afaik its a once off up-front cost).
(* Running on Linux doesn't mean it magically works, but the Tilera PCIe boards can run Linux, so it at least makes porting a bit easier)
But that doesn't change the facts: Intel DOES have competition to their high-end market from GPUs and the likes of Tilera. I can see the newer ARM processors starting to challenge their midrange line and obviously with Atom too.
Just today, I was thinking about getting an ARM Cortex A8 powered BeagleBoard xM as adesktop linux box replacement.
it would be interesting to know the lifetime cost(including power, cooling and space) relative to the x86.
But Intel is not falling for that one. They have made the Atom, a slow, cheap, low-power chip that competes directly with ARM. That's likely a wise move, but Intel now has the problem that low-margin chips are still bad business. They have to have their expensive best-in-industry fabs make low-margin Atoms, when they would much, much rather have them make expensive Xeons.
At one point they made a deal to have Atoms manufactured at TSMC, which would have helped a lot with this problem, but apparently that deal didn't work out. Even if it had worked out, the Atom would no longer have the process advantage, and then backwards compatibility would be the only advantage for x86. With Windows becoming less and less relevant, that's a big problem, considering the technical advantages ARM has over x86.
So fundamentally, Intel has a problem that CPUs are becoming commoditized, which means they will either have to take much lower margins or retreat to the high end. Both scenarios are unpleasant for them.